facebookresearch/PoincareMaps

The need to understand cell developmental processes has spawned a plethora of computational methods for discovering hierarchies from scRNAseq data. However, existing techniques are based on Euclidean geometry which is not an optimal choice for modeling complex cell trajectories with multiple branches. To overcome this fundamental representation issue we propose Poincaré maps, a method harnessing the power of hyperbolic geometry into the realm of single-cell data analysis.

PoincareMaps

Poincare maps recover continuous hierarchies in single-cell data.

POC: Anna Klimovskaia (klanna@fb.com)

Dependecies

python3.7 anaconda (sklearn, numpy, pandas, scipy) seaborn

Pytorch (pytorch 1.7.1): https://pytorch.org/get-started/locally/

To replicate our experiments

Embedding

python main.py --dset ToggleSwitch       --batchsize -1 --cuda 1 --knn 15 --gamma 2.0 --sigma 1.0 --pca 0  --root root
python main.py --dset MyeloidProgenitors --batchsize -1 --cuda 1 --knn 30 --gamma 2.0 --sigma 2.0 --pca 0  --root root
python main.py --dset krumsiek11_blobs   --batchsize -1 --cuda 1 --knn 30 --gamma 2.0 --sigma 1.0 --pca 20 --root root

python main.py --dset Olsson   			 --batchsize -1 --cuda 1 --knn 15 --gamma 2.0 --sigma 1.0 --pca 20 --root HSPC-1
python main.py --dset Paul               --batchsize -1 --cuda 1 --knn 15 --gamma 2.0 --sigma 1.0 --pca 20 --root root
python main.py --dset Moignard2015       --batchsize -1 --cuda 1 --knn 30 --gamma 1.0 --sigma 2.0 --pca 0  --root PS
python main.py --dset Planaria           --batchsize -1 --cuda 1 --knn 15 --gamma 2.0 --sigma 2.0 --pca 0 --root neoblast\ 1

python main.py --dset MyeloidProgenitors --batchsize -1 --cuda 1 --knn 30 --gamma 2.0 --sigma 2.0 --pca 0  --root root
python main.py --dset Olsson   			 --batchsize -1 --cuda 1 --knn 15 --gamma 2.0 --sigma 1.0 --pca 20 --root HSPC-1
python main.py --dset Planaria           --batchsize -1 --cuda 1 --knn 15 --gamma 2.0 --sigma 2.0 --pca 0 --root neoblast\ 1

Prediction

python decoder.py --dset Planaria --cuda 1 --method poincare
python decoder.py --dset Planaria --cuda 1 --method UMAP
python decoder.py --dset Planaria --cuda 1 --method ForceAtlas2

Structure of the repository

Folder datasets contains datasets used in the study.

Folder results contains Poincaré map coordinates.

Folder decoder contains weights of the pretrained decoder network.

Folder predictions contains coordinates of sampled (interpolated) points.

Folder benchmarks contains visualization of benchmark embeddings.

License

PoincareMaps is Attribution-NonCommercial 4.0 International licensed, as found in the LICENSE file.